与颜色空间选择无关的肤色检测方法

Skin Color Detection without Considering the Choice of Color Space

针对当前肤色检测方法中如何选择合适的颜色空间和高准确率伴随高误报率的问题,提出一种基于柔性神经树的肤色检测方法.该方法随机生成不同结构的柔性神经树,在评估函数的约束下使用基于语法引导的遗传算法进行结构优化,同时使用粒子群优化算法对参数进行优化,最终获得最优肤色模型.优化过程保留了结构简单和具有较好性能的个体,使得训练得到的肤色模型保留了颜色空间中贡献较大的特征分量.在Compaq和ECU数据库上的实验结果表明,文中方法分别获得了92.27%和90.79%的准确率以及6.16%和9.71%的误报率,优于其他主流肤色检测方法,且对复杂环境下肤色细节的检测有较好的效果.

There are problems in skin detection methods： choice of suitable color space and methods with high accuracy usually accompanies high false positive. To address these problems, a novel skin detection method is proposed based on flexible neural tree. A set of flexible neural trees with different structures are generated firstly, which will be optimized in structure using grammar guided genetic programming and in parameters using particle swarm optimization algorithm with regulation of evaluation function simultaneously. Individuals with simplicity structure and similar accuracy models were selected and reserved in the training stage, i.e., the components having greater contributions were reserved in the trained model. Experiment on Compaq and ECU data sets shows the proposed method achieves 92.27% accuracy with false positive of 6.16% on Compaq set, and 90.79% accuracy with false positive of 9.71% on ECU set, which outperforms state of the art methods. The experiment also illustrates good detailed skin detection in challenging scenarios.